Molecular design in drug discovery: a comprehensive review of deep generative models Briefings in Bioinformatics

Table Of Content

• Molecular and Computational Biology
• True shape of lithium revealed for the first time in UCLA research
• Optimization strategy for molecular design
• 5. Inverse Molecular Design
• Quantum Trailblazers: NarangLab’s Pursuit
• Source Data Fig. 3
• Evolutionary design of molecules based on deep learning and a genetic algorithm

In recent years, some improvement on the network architectures like long-short term memory (LSTM) [57] and gated recurrent unit (GRU) [58] have been proposed due to the difficult training of RNNs. LSTM, adding the memory cell that replaces conventional units, solves difficulties with training encountered by RNNs. And the simplicity of GRU is more suitable for building larger networks due to the smaller amount of parameters. Currently, CAMD workflows are generally built and trained with a specific goal in mind. Such workflows need to be re-configured and re-trained to work for different objectives in therapeutic design and discovery. It would be particularly very helpful for the domains where a relatively small amount of data exist.

Molecular and Computational Biology

Due to the unstable training of GAN, some variants were proposed like wassertein GAN (WGAN) [52]. WGAN incorporates the Earth-Mover (EM) distance, which reflects the minimum cost under optimal planning to get a smoother gradient. WGAN not only alleviates the problem of unstable training, but also evaluates reliably generative models to avoid mode collapse. Visualization of the latent representations via t-SNE obtained with the trained energy-based model for the molecules in the training set as well as the generated molecules with restrictions on the QED property is shown. 5b, as the number of phases increases, the number of new molecules with S1 values that are relatively lower than those in the previous phase increases. After the first phase, the number of newly generated molecules with S1 values lower than 1.77 eV is 12.

True shape of lithium revealed for the first time in UCLA research

QbD for Small-Molecule Continuous Process Development - Pharmaceutical Technology Magazine

QbD for Small-Molecule Continuous Process Development.

Posted: Fri, 02 Feb 2024 08:00:00 GMT [source]

The graph-based deep learning model MGM is able to generate a few molecules that satisfy corresponding targets but is accompanied by the generation of other noncomplying structures in a significant proportion. As evident from the efficacy metrics reported in Table 2, a deviation as high as 72.3% from the mean target property requirements can be observed with the baseline deep learning methods for molecular generation. In contrast to the graph and autoencoder-based baselines, GBGA limits the search of molecules candidates within a narrow domain for the QED property targets but exhibits wider exploration for the LogP property targets. On the other hand, the proposed QC-based approach is able to generate molecules that exhibit target properties efficiently with observed zero violations of the required target property constraints.

Optimization strategy for molecular design

Governed by the proposed optimization procedure, the surrogate model is sequentially refined to explore the chemical space for identifying molecules that satisfy the desired property requirements and structural constraints. Since the sequence representation of SMILES, the analogy of natural language processing tasks and molecular generation is feasible [27]. For RNNs, the features obtained from large molecular datasets can be transferred to produce molecules with activity on demand in small ones, so that Segler et al. [59] generated focused molecule libraries by retraining the model (refer Figure 2.3). Sampling from the large-scale datasets ensured the diversity of molecules and fine-turning increased the focused properties.

Having such a CAMD infrastructure, algorithm and software stack would speedup end-to-end antiviral lead design and optimization for any future pandemics, such as COVID-19. In the decreasing direction, the maximum rates of change in S1 are slightly lower than those in the increasing direction, which may be caused by the S1 distribution of the training data. 2, in the case of 50,000 samples of training data, the S1 distribution is skewed and is higher than the median S1 value of the seed molecules, i.e., 4.0 eV. The average S1 values are 4.4, 4.3, and 4.4 eV for 10,000, 30,000, and 50,000 samples of training data, respectively. Owing to the characteristics of the training data, S1 is more likely to change its value in the increasing direction.

All quantum chemical calculations were performed with the Gaussian 09 program suite38. The molecular geometries were optimized by density functional theory (DFT) using the hybrid B3LYP functional and all-electron 6-31G basis sets. A single-point time-dependent DFT calculation was performed with this geometry to calculate the vertical excitation energies to the lowest singlet state (S1). The researchers solve this issue by building a model that runs directly on molecular graphs, instead of SMILES strings, which can be modified more efficiently and accurately.

The distribution of the proportion of molecular candidates satisfying target requirements obtained with the energy-based models trained with both CD learning and QC-assisted learning are plotted for c QED property targets and d LogP property targets. The same set of reference molecules is used as the initial starting point for optimizing molecules with both models for a fair comparison. Generative networks based on RNNs model the graph generation as a sequential process and make auto-regressive decisions while they generate graphs. GraphNet [81], the first RNNs-based model on arbitrary graph, was on the framework of the message-passing neural networks (MPNN) [82]. The essence of GraphNet was to add a new atom or bond into the existing graph.

Data Availability Statement

Moreover, the experimental process involves a series of steps, each requiring several correlated parameters that need to be tuned [2,3], which is a daunting task, as each parameter set conventionally demands individual experiments. This has slowed down the discovery of high-impact small molecules and/or materials, in some case by decades, with possible implications for diverse fields, such as in energy storage, electronics, catalysis, drug discovery, etc. Molecular modeling methods used to study protein–ligand interactions including molecular docking simulations, molecular mechanics methods, hybrid Quantum Mechanics/Molecular Mechanics simulations, and deep learning models for the activity and affinity prediction. Figure 2 shows the average rate of change of S1 for the 50 seed molecules when the number of training data samples increases from 10,000 to 50,000.

Evolutionary design of molecules based on deep learning and a genetic algorithm

The Virtual Model Kit has been a source of inspiration for the birth of this project. A promising drug, for example, might prove to have low toxicity in general, but one disturbing side effect. The Hansch equation suggests how the molecule should be modified to minimize the adverse properties.

2, the S1, HOMO, and LUMO distributions for 10,000 and 30,000 samples of training data are similar to those of the 50,000 samples of training data. To obtain the property prediction function f(∙), a five-layer DNN was built with 250 hidden units in each layer to identify the nonlinear relationship between molecular structures and their properties36. This research focuses on the characterization of a simple organic molecule incorporating biphenyl, methacrylate, trimethylsilyl acetylene, and liquid crystal compounds.

Accordingly, the RNN and DNN models were trained with a chemical library that comprises 10,000 to 100,000 molecules (with molecular weights between 200 and 600 g/mol) randomly sampled from the PubChem database32. Each molecule was labeled with the excitation energy (S1), molecular orbital energies (highest occupied molecular orbital (HOMO), and lowest unoccupied molecular orbital (LUMO)) by the DFT calculation. As the amount of training data increases, the performance of the RNN and DNN models improves accordingly, as summarized in Table 1. For each data split, we trained the prediction model using the training set and evaluated its prediction performance on the test set. The validity of RNN decoding, which refers to the proportion of chemically valid molecules, was assessed during the RDKit inspection step.

Genetic algorithms (GAs) have also been used for generating molecules while optimizing their properties [103,104,105,106]. GA-based models suffer from stagnation while being trapped in at the regions of local optima [107]. One notable work alleviating these problems is by Nigam et al. [56], where they hybridize a GA and a deep neural network to generate diverse molecules while outperforming related models in optimization.

Moreover, deep neural network (DNN) models31, aided by quantum chemical calculations, are used to evaluate the evolved molecules with more complex criteria. Since SMILES is regarded as string of texts, a large number of models in natural language processing are able to be extended to the field of de novo molecular design. In future research, for example, we can consider the molecular generation for desired properties as a translation, which can translate from the specific target language (protein sequence) to the SMILES language. Notably, despite the surge of SMILES-based models in recent years, there are still some burning problems. Not only is it facing the issues of validity, but the unstructured nature of SMILES makes two similar molecules be quite different with a high probability. And it is expensive to force the validity constraint to incorporate into the decoders, which requires for designing a novel representation with more structural information.